Import AI 453: Breaking AI agents; MirrorCode; and ten views on gradual disempowerment

# AI Agents Still Have Major Security Gaps Researchers have found that AI agents—systems designed to act independently and make decisions—can be tricked or broken in ways that weren't previously understood, raising questions about whether we can safely deploy them in the real world. The article also explores whether major AI breakthroughs might feel as transformative to us as the discovery of fire did to our ancestors, suggesting we may not fully recognize how fundamental some changes are until much later.